Educational Management Strategies and Learning Effectiveness Enhancement of Information Technology Integration in Information Technology Reform of English Education

The K-means clustering algorithm belongs to unsupervised learning in machine learning. The difference between unsupervised learning and supervised learning is the presence or absence of learning labels. K-means clustering divides object into k clusters based on their attributes to satisfy the higher similarity of the objects in the same cluster, and the lower similarity of the objects in the different clusters. K-means clustering uses iterative method to find the center of each cluster. That is, each data sample is divided into the cluster of the nearest centroid, and then the center of each cluster is updated continuously until the center of each cluster no longer moves [33].

When calculating the center of each cluster, usually we need to calculate the distance between the values of each attribute, which can be simply taken as the average of the X/Y-coordinates of each point, and in general the Minkowski distance formula is used, as shown in (1): (1)d(Xi,Xj)=(∑i=1n|Xik−Xjk|p)1p$$d\left( {{X_i},{X_j}} \right) = {\left( {\sum\limits_{i = 1}^n {{{\left| {{X_{ik}} - {X_{jk}}} \right|}^p}} } \right)^{\frac{1}{p}}}$$

When p = 1, it is called the Manhattan distance, also known as the absolute distance and the city block distance: (2)d(Xi,Xj)=∑i=1n|Xik−Xjk|$$d\left( {{X_i},{X_j}} \right) = \sum\limits_{i = 1}^n {\left| {{X_{ik}} - {X_{jk}}} \right|}$$

When p = 2, it is called the Euclidean distance and is the most widely used distance formula: (3)d(Xi,Xj)=(∑i=1n|Xik−Xjk|2)12$$d\left( {{X_i},{X_j}} \right) = {\left( {\sum\limits_{i = 1}^n {{{\left| {{X_{ik}} - {X_{jk}}} \right|}^2}} } \right)^{\frac{1}{2}}}$$

In addition to this, the commonly used distances are the Mahalanobis distance, Lang’s distance, and so on.

The main flow of K-means algorithm is shown in Figure 1. Using K-means algorithm to cluster analysis of students’ performance in various subjects, the selection of k value is very important, if the k value is not selected well, the clustering results will be very poor, therefore, it is necessary to run several times to select the initial cluster centers according to certain criteria.

Figure 1.

K-means clustering algorithm process

In this flowchart, the number of clusters k is first determined based on the input dataset, which is the number of samples n, and the clustering center is initially determined. The commonly used Euclidean distance is chosen as the formula for calculating the distance to determine the distance from each object to the clustering center d, and divided into the nearest clusters, using an iterative method to continuously update the centroid until the centroid position is no longer changed, and output the clustering results.

Particle Swarm Algorithm (PSO) relies on the Boid (bird-bid) model to find the optimal value, following the rules of conflict avoidance, speed matching, and group centering. The essence is that the particles keep doing directional variable-speed motion in space to find the next position through their own memory and the communication of the group to find the optimal solution [34].

The particle swarm algorithm is expressed as follows: n represents the total number of particles in the swarm, the position of the ind particle in the D-dimensional space is represented by X_i, the speed of change of the particle is represented by V_i, the best position searched by the particle in the target space is represented by P_i, and the best position searched by the whole particle population is represented by Pg. The particle swarm algorithm has to update its own velocity and position during each iteration with the following equations: (4)Vidk+1=ωVidk+c1r1(Pidk−Xidk)+c2r2(Pgdk−Xidk)$$V_{id}^{k + 1} = \omega V_{id}^k + {c_1}{r_1}\left( {P_{id}^k - X_{id}^k} \right) + {c_2}{r_2}\left( {P_{gd}^k - X_{id}^k} \right)$$ (5)Xidk+1=Xidk+Vidk+1$$X_{id}^{k + 1} = X_{id}^k + V_{id}^{k + 1}$$

c₁c₂ is the acceleration coefficient, r₁r₂ is the random number in [0, 1], ω is the inertia factor, which is a non-negative number, and the larger the value of ω, the larger the variation of the particles, and the local optimal solution will not be found, but the global search ability is stronger at this time. After experimental testing, the value of ω is 0.6-0.75, and the number of particles is 32, it is easy to find the global optimal solution.

1)

Model Design

The publicly available student achievement dataset on UCI was selected, which contains 29 attributes related to students’ learning behaviors such as gender, age, parents’ work situation, parents’ education level, students’ study time, after-school activity time, recreation time, etc. The preprocessed data were then divided into training set and test set according to the ratio of 7:3, and the XGBoost algorithm model was introduced on the basis of the Particle swarm algorithm, so as to be able to carry out the optimal parameter selection in classification, thus improving the accuracy of the classification algorithm.The whole process of the PSO-XGBoost-based early warning model of students’ academic performance is shown in Fig. 2.The XGBoost has a good characteristic of preventing overfitting and high computational efficiency.After selecting the feature segmentation points as leaf nodes at each layer, it makes the gain value of the tree become larger, which means that these features have been segmented a number of times. That is to say, the more times these features are segmented, the greater the gain of the whole tree and the corresponding features become more important. In feature selection, the PSO algorithm is used to find the global optimal solution, and model design based on the optimal features is aimed at combining the advantages of the two algorithms, eliminating the influence of discrete data, preventing overfitting, and improving the accuracy of the model.

Figure 2.

Academic performance early warning model

where F denotes the function space containing all the trees, and L(yi,y^i(m))$$L\left( {{y_i},\hat y_i^{(m)}} \right)$$ is the loss function that measures the difference between the predicted value yim$$y_i^m$$ and the target value y_i, i.e., the difference between the predicted value and the true value: (7)L(yi,yi(m))=(yi−y^i(m))2$$L\left( {{y_i},y_i^{(m)}} \right) = {\left( {{y_i} - \hat y_i^{(m)}} \right)^2}$$

Ω(fm)$$\Omega \left( {{f_m}} \right)$$ is the regularization term, defined as shown in Eq. (8): (8)Ω(fm)=γT+12λ∑j=1Tωj2$$\Omega \left( {{f_m}} \right) = \gamma T + \frac{1}{2}\lambda \sum\limits_{j = 1}^T {\omega _j^2}$$

where the number of leaf nodes is denoted by T, the regularization parameter is denoted by λ, γ is the learning rate, and ω_j is the predicted value of the jth leaf node. Using the residuals to fit the final value in multiple iterations can significantly improve the accuracy of the algorithm.

The expression of its prediction function is shown in Equation (9): (9)y^i(t)=∑K=1Kfk(xi)=yi(t−1)+ft(xi)$$\hat y_i^{(t)} = \sum\limits_{K = 1}^K {{f_k}} \left( {{x_i}} \right) = y_i^{(t - 1)} + {f_t}\left( {{x_i}} \right)$$

where f(x) = ω_j(x), ω_j(x) are used as scores for leaf node j and f_k(x) is the kth regression tree.

Training error: f_k(x) denotes the weight of the ith sample falling in the leaf in the kth tree, θ = f₁, f₂, f₃, …, f_k: (10)L(θ)=∑i=1nl(yi,y^i)=∑i=1nl(yi,∑k=1kfk(xi))$$L(\theta ) = \sum\limits_{i = 1}^n l \left( {{y_i},{{\hat y}_i}} \right) = \sum\limits_{i = 1}^n l \left( {{y_i},\sum\limits_{k = 1}^k {{f_k}} \left( {{x_i}} \right)} \right)$$

Model Complexity: Ω(f) indicates the model complexity: (11)Ω(θ)=∑k=1kΩ(fk)$$\Omega (\theta ) = \sum\limits_{k = 1}^k \Omega \left( {{f_k}} \right)$$

Academic Alert modeling is a problem that involves classification of multiple categories; therefore, we can assess the classification effectiveness of the model by using evaluation metrics. Evaluation metrics are important tools used to measure the quality and effectiveness of model training, and different scenarios require the use of different evaluation metrics. Among the common evaluation metrics, they include precision rate, recall rate, F1 score (F1−score)$$\left( {F1 - score} \right)$$, and so on. These evaluation metrics can provide us with feedback about the model’s performance and performance in different aspects, thus helping us to better optimize and improve the model.

Precision rate indicates the proportion of samples predicted to be positive cases that are actually positive cases. Its calculation formula is: (12)Precision=TPTP+FP$$\Pr ecision = \frac{{TP}}{{TP + FP}}$$

Recall indicates the proportion of positive examples in the original sample that are predicted correctly. In unbalanced datasets, the recall of a few classes tends to receive greater attention. The recall rate is calculated as: (13)Recall=TPTP+FN$$\operatorname{Re} call = \frac{{TP}}{{TP + FN}}$$

F1 Score is a metric used to measure the precision of a binary classification model. Precision and recall are contradictory metrics; in general, recall tends to be low when precision is high. When the recall is high, the precision tends to be low. The F1-score takes precision and recall into account and is a weighted average of them. Its calculation formula is: (14)F1−score=2*Precision*RecallPrecision+Recall=2TP2TP+FP+FN$$F1 - score = \frac{{2*{\text{ }}\Pr ecision*\operatorname{Re} call}}{{\Pr ecision + \operatorname{Re} call}} = \frac{{2TP}}{{2TP + FP + FN}}$$

In this paper, 300 students majoring in English in college A were selected as the research object, and four items of student learning behavior data provided by the Learning Pass platform were extracted from the statistical data of the Learning Pass platform, namely, the video viewing time, the number of visits, the percentage of completion of the task points and the quiz scores. In the process of extracting the learning behavior data, in addition to retaining the four learning behavior data also retained the student’s school number and name data, which is convenient to understand the student’s information after classifying the students at a later stage. After removing the data with zero logins, a new set of learning data was obtained after extracting and cleaning the learning data, and a total of 5 students’ learning behavior data were cleaned and 295 learning behavior data records were retained.

Due to the different orders of magnitude of the extracted learning behavior data, this paper standardizes the extracted data. In this paper, the standard deviation standardization method (Z-Score) is used to standardize the learning behavior data.

After the necessary pre-processing of the learners’ behavioral data, the article adopts the K-Means clustering method as a research tool. While applying the K-Means clustering algorithm, the first and foremost task is to determine the appropriate k value. For this purpose, this paper evaluates the effect of different k values on the clustering effect with the help of SSE (Sum of Squared Errors) metric. By analyzing the relationship between the number of clusters k and SSE, a clear “inflection point” is found, and based on this, the optimal number of clusters k is selected. The SSEs corresponding to different values of k are shown in Fig. 3. When the number of clusters k is taken as 3, the corresponding SSE value reaches 1.11. Based on this finding, the number of clusters is chosen to be 3 in this study.

Figure 3.

The different k values correspond to SSE

Next, data cleaning was performed on the characteristic attributes of learners’ learning actions, followed by clustering using the K-means method. The distribution of clusters is shown in Figure 4. According to the distribution of clusters of learners’ learning actions in the figure, we can analyze:

Figure 4.

Distribution of clustering

Cluster 3 represents 66.5% of the total. Learners in this cluster show a high degree of learning engagement and academic achievement. Characteristics include higher frequency of course visits, higher assignment submission rates, and higher test scores. It can be concluded that these students have good study habits and strong learning abilities.

Cluster 1 is responsible for 9.8% of the total. Learners in this cluster show lower learning engagement and learning achievement. Their characteristics include lower frequency of course visits, assignment submission rates, and test scores. It can be inferred that these learners may lack motivation or face additional barriers to learning.

Cluster 2 makes up 23.7% of the total population. Learners in this cluster exhibit moderate levels of academic engagement and academic achievement. Characteristics include a moderate frequency of course visits and a high rate of assignment submission. Compared to Cluster 1, these learners may have experienced some difficulty in the learning process, but still possess some learning ability.

The change in the number of English academic warning is shown in Figure 8. Before the implementation of the early warning system (the first semester of 2021-2022) the number of early warnings per semester in English majors in college A was high. After the implementation of the early warning system, the number of early warnings in English majors in college A shows a significant downward trend, and by the second semester of 2022-2023, the number of academic early warnings has dropped to 52. It can be seen that academic early warning plays a timely role in supervising most students, especially students in difficulty, so that students pay attention to their studies at any time, avoid and reduce failing classes, and reduce the number of academic early warning as a whole. The decrease in the number of academic alerts leads to fewer academic problems that are more difficult for counselors and academic advisors to communicate with students and parents, and reduces the workload of teachers and teaching administrators. The decline in the number of students on academic warning proves that the number of students who have difficulties in their studies is gradually decreasing, and their grades have changed from failing to passing and from moderate to excellent, and the academic warning mechanism has gradually become an important part of teaching quality assurance.

Figure 8.

The number of academic early warning changes

The change in the number of struggling English students after the pilot run of the academic alert system for one and a half years is shown in Figure 9. It was reduced from 87 to 14 before the academic warning system was introduced. It shows that it has played a good role in monitoring and early warning for the academics of students who are in difficulties, especially. With the gradual implementation of the early warning system and the timely information warning in place, students began to face up to every warning information they received, through timely attention, can check the shortcomings and make up for the shortcomings, and gradually the coursework took off, no longer become a frequent flunking out of the class, and the students gradually recognized the role played by the academic early warning. In just one and a half years, the number of students in difficulty has dropped dramatically.

Figure 9.

Change the number of students who are in difficulty